Public | Automated Build

Last pushed: a year ago
Short Description
Optimized versions of the Tensorflow docker images.
Full Description

Tensorflow with AVX2 and CUDA

This is a Docker image for running Tensorflow 0.9 with support for AVX2 and
CUDA. In my testing, this is 30-100x faster than the default Tensorflow image
(gcr.io/tensorflow/tensorflow).

The Tensorflow binaries included here are built from upstream 0.9 with
--copt=-mavx2. The container does not actually include the dependencies
needed to build Tensorflow, because they're kind of scary. I've included
Tensorflow's license as tensorflow.LICENSE.

Currently, this only works well on a Linux host using the
nvidia-docker wrapper.

Like the stock Tensorflow image, this container will come up running Jupyter on
port 8888.

To build:

docker build github.com/cbiffle/docker-tensorflow:avx2-cuda

To run (from Docker Hub):

nvidia-docker run -it -p 8888:8888 cbiffle/docker-tensorflow:avx2-cuda

Or using an external notebook directory (which I recommend, as it makes the
container ephemeral):

nvidia-docker run -it -p 8888:8888 -v /path/to/project:/notebooks \
    cbiffle/docker-tensorflow:avx2-cuda
Docker Pull Command
Owner
cbiffle
Source Repository

Comments (0)